PATTERN-CLASSIFICATION FROM DISTORTED SAMPLE

In nonparametric pattern classification the Optimal (Bayesian) decision on the category of the observed vector is designed from a long training sequence, that is, independent pairs of observations and corresponding labels. In many practical situations, however, due to feature extraction, quantization, or noise, the observed vector and the training sequence may be distorted. In this paper we show how asymptotically optimal decisions can be derived from distorted training or made from slightly distorted observation.